spectrometerHBH opened a new pull request, #19895:
URL: https://github.com/apache/tvm/pull/19895
This adds an in-kernel profiling page to the TIRx native-basics CUDA section,
documenting the existing `tvm.tirx.bench.CudaProfiler`.
The page covers:
- a minimal load / compute / store example using `start` / `end` / `finalize`
markers and a user-supplied `uint64` buffer;
- decoding the record buffer on the host and exporting a Perfetto trace via
`export_to_perfetto_trace`;
- the record/tag encoding and the device code each call lowers to (a
`%globaltimer` read, a leader-only global store, and a block fence);
- usage notes: one leader per `(block, group)`, buffer sizing, the 32-bit
`%globaltimer` wrap, and the per-region cost.
The example is tested end-to-end on a CUDA GPU (B200, sm_100). It is wired
into
the `native_basics.rst` toctree after "Compiling and inspecting". The
FlashAttention-4 timeline screenshot is served from `tlc-pack/web-data`
(`images/tirx/tirx_cudaprofiler_fa4.png`), matching the other tirx doc
figures.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
